This repository implements the paper "Reference based Sketch Image Colorization using Augmented-Self Reference and Dense Semantic Correspondence" published in CVPR2020.
- python3.6+
- pytorch 1.6.0
- others.
training a model
python3 main.py --config config.yml
testing a model
Not implmented yet
In this implementation, the triplet loss function is meaningless. It always show zeros for scaled dot product and l2 norm distance, if I am wrong, please make issue. Without the triplet loss, we can obtain good results. Even if a model is trained only 2 epochs, the model shows meaningful results.
- tps_transform : https://github.com/cheind/py-thin-plate-spline
- spectral normalization : https://github.com/christiancosgrove/pytorch-spectral-normalization-gan/blob/master/spectral_normalization.py
- unet : https://github.com/milesial/Pytorch-UNet
- dataset : https://www.kaggle.com/ktaebum/anime-sketch-colorization-pair